From a queuing theory perspective, what really happens when capacity is added depends on if additional transaction processors (think more CPU cores) are implemented or the transaction processors are faster (think faster CPUs)-or if we're lucky, both. If the transaction processors are faster, the service time decreases with the same effect as with tuning. We can expect a new response time curve similar to the one shown in Figure 9-17 to take effect. However, if we add transaction processors with no change to service time, the response-time curve does not shift down. But because there are more transaction processors available, as a whole, they can process more transactions per unit of time, which shifts the curve to the right, allowing for an increase in the arrival rate before queuing sets in.

Figure 9-18 graphically shows how adding more transaction processors can affect a system. If the bottleneck is IO, then the same general effect occurs when adding IO devices. Starting at point A, the performance is unacceptable and highly variable. By implementing addition transaction processors, the response time decreases (that is, improves), and the system is operating at point B. However, now the administrators have a choice. By controlling the workload (the arrival rate), they can allow more work to flow through the system without affecting response time. Point C shows this negligible effect on response time by allowing the arrival rate to increase. So, by adding more capacity (either more and/or faster transaction processors), the performance analyst once again has several options: decreased response time, increased workload, or a managed combination of both.

Figure 9-18. Shown is the response-time effect of increasing capacity by adding transaction processors (for example, CPU cores). By adding CPU cores, a new response-time curve takes effect (dotted line). The response time drops from point A to point B. By controlling the workload, performance can remain at point B, or by allowing the workload to increase to point C, the system can still maintain both improved response time and an increased workload.